Against the background of intelligent transformation and high-quality development in the shipbuilding industry, large models represented by large language models (LLMs) and multimodal large models have gradually become core technologies driving the intelligent upgrading of the entire ship life cycle. This paper systematically reviews the research progress of integrating large model technologies with the shipbuilding industry, and clarifies the technical principles, capability boundaries and existing limitations of large models in engineering applications. By summarizing the development route of artificial intelligence and the underlying mathematical logic of large models with Transformer as the core architecture, this study analyzes the key capabilities of large models in natural language interaction, multi-step reasoning, multimodal understanding, few-shot learning and code generation, as well as inherent defects such as hallucinations, weak causal reasoning and poor real-time performance.According to the engineering characteristics and safety requirements of the shipbuilding industry, this paper proposes three basic application principles for large models: 1) introducing ship industry knowledge as constraints to improve output reliability; 2) taking large models as planning engines to realize the decomposition of complex tasks and the collaboration of professional tools; 3) and building multi-agent systems with large models as the decision-making center to support autonomous decision-making and closed-loop execution. Guided by these principles, this paper sorts out and evaluates the application cases of large models in four typical scenarios: ship design, ship manufacturing, operation and maintenance, and navigation decision-making. It is found that the knowledge-constrained scheme is suitable for standardized scenarios such as rule compliance and design consulting; the planning engine framework is efficient in complex task decomposition; and the agent system shows unique advantages in autonomous decision-making and multi-ship collaboration.Furthermore, this paper discusses the technical bottlenecks restricting the large-scale engineering implementation of large models in the shipbuilding industry, including data quality and governance, real-time response for navigation control, decision interpretability and safety verification. On this basis, a step-by-step integration idea of "technology alignment–scenario adaptation–system implementation" is put forward to promote the reliable landing of large models in the whole life cycle of ships. Finally, the future development trends are prospected from the aspects of multimodal fusion, causal reasoning enhancement, human-in-the-loop reinforcement learning and digital twin-driven simulation verification. The conclusions and methods of this paper can also provide theoretical references and practical guidance for the intelligent transformation of other complex industrial fields such as aerospace, astronautics, energy and electrics.
- Article type
- Year
- Co-author
The emergence of large language models (LLMs) is profoundly transforming several fields, including education. Currently, LLMs have been successfully utilized to assist in general education, liberal arts, and computer science teaching fields. However, the application of LLMs in power electronics experimental teaching still faces obstacles, challenged by issues such as inaccurate circuit topology recognition and omission of key circuit information, including circuit component connections. Therefore, this paper proposes an innovative artificial intelligence method for power electronics experimental teaching, leveraging the synergistic combination of intelligent plugins and LLMs to achieve accurate circuit topology recognition, information extraction, knowledge verification, and content generation. The primary goal is to overcome existing limitations and enhance the efficiency and quality of experimental teaching.
The proposed method first trains a deep learning model, namely YOLO, to accurately identify the components, nodes, and positions of the circuit topology. Then, it designs an algorithm to translate this information into circuit netlist text, which is recognizable by LLMs. This netlist, along with a series of carefully crafted prompts, is presented to the LLM to guide it in completing various tasks, including netlist verification, circuit analysis, and document generation. The proposed method encapsulates circuit topology recognition and netlist generation into “intelligent plugins,” compensating for the inherent shortcomings of the LLM. This paper also designs a series of prompts to guide the LLM in effectively performing its designated teaching functions, such as generating experimental instructions for teachers and generating preview Q&A sets for students. To ensure cooperation among the intelligent plugins, prompts, and LLM, the workflow of the proposed method is also presented. In addition, a software tool named “PE Exp. Agent,” which can be integrated with various LLMs through an API interface, vividly demonstrates the aforementioned design philosophy.
This software tool is applied to several stages of power electronics experimental teaching. The intelligent plugins accurately identify the circuit topology from the published paper and correctly generate the corresponding netlists. By applying different prompts, the LLM generates different content. In particular, different LLMs are called on for extensive testing, and the results showed that, although differences exist in the style and expression among different LLMs, they always remain consistent with the requirements set by the prompts. Thanks to the natural language understanding capabilities of LLMs, any dissatisfaction can be quickly addressed and corrected through interaction with the LLM. Therefore, the proposed method considerably reduces the workload of power electronics experimental teaching personnel and improves the timeliness and quality of experimental teaching.
The integration of intelligent plugins and LLMs provides a novel and effective method for the design of power electronics experimental teaching. It not only lightens the burden on teachers but also improves the overall quality of teaching. By developing software tools similar to “PE Exp. Agent,” the efficiency and effectiveness of experimental teaching can be effectively promoted. With modifications to the plugins and prompts, this method can also adapt to the needs of more teaching tasks. The method proposed in this paper shows potential as a new productivity tool for the design of power electronics experimental teaching.
“Power Electronics Technology” is a highly applied and practical discipline, usually equipped with relevant power electronics experimental teaching segments. However, the traditional customized, standardized and basic experimental platform has poor teaching effectiveness, and it is difficult to replace experimental cases. To address the above issues, this article proposes a new power electronics experimental teaching platform based on the concept of module standardization and design automation. Its advantages include: ① It can be quickly assembled while giving students enough testing freedom and innovation space; ② Compact and elegant layout, excellent performance, and good teaching demonstration effect; ③ Equipped with automatic aided design software, which greatly shortens the course development cycle. These advantages fully align with the concepts of practicality, innovation, and engineering cognition of “New Engineering”, and can significantly improve the quality of experimental teaching.
京公网安备11010802044758号